Velocity Field: An Informative Traveling Cost Representation for Trajectory Planning

TL;DR

A local map representation method that provides heading and velocity priors for trajectory planning tasks, simplifying the planning process in complex urban driving scenarios and investigating local map representation forms for planning performance on a real-world dataset.

Abstract

Trajectory planning involves generating a series of space points to be followed in the near future. However, due to the complex and uncertain nature of the driving environment, it is impractical for autonomous vehicles~(AVs) to exhaustively design planning rules for optimizing future trajectories. To address this issue, we propose a local map representation method called Velocity Field. This approach provides heading and velocity priors for trajectory planning tasks, simplifying the planning process in complex urban driving. The heading and velocity priors can be learned from demonstrations of human drivers using our proposed loss. Additionally, we developed an iterative sampling-based planner to train and compare the differences between local map representations. We investigated local map representation forms for planning performance on a real-world dataset. Compared to learned rasterized cost maps, our method demonstrated greater reliability and computational efficiency.

Figures (3)

• Figure 1: This figure illustrates three methods for calculating the cost of travel at a given time step. The red car represents the ego vehicle, the solid white line on the far right depicts the stop line for the traffic light, and the traffic lights in each scenario are in the red light state. (a) distances between the ego vehicle and other vehicles and traffic signals are measured. Manually defined rules are then implemented to determine the car-following speed and lane change maneuvers. (b) represents a class of map representations that provide the driving cost for each point in the driving context. It guides the vehicle trajectory through as many low-driving cost areas as possible. (c) uses the velocity vector to guide the vehicle's driving trajectory. For example, it instructs vehicles to slow down and stop before the stop line around stop signs and to drive at the same speed as other cars.
• Figure 2: The diagram presents our proposed framework for velocity field generation and planning modules, along with their training process. The green arrows represent the transmission of latent variables. Reference lines, map elements, ego and agent motion, and geometry are encoded using MLP/GRU modules. Global attention is employed to model the interaction between objects. The multi-modal MLP decoder generates initial guesses for the ego vehicle and predictions for other agents. The orange arrows indicate the relationship between loss and each module: expert demonstrations provide supervisory signals to model parameters through loss functions, such as distance measures, selection probabilities, and velocity targets. The black arrows represent the planning process: the iterative trajectory optimizer samples around the initial guess; calls the inquiry function to obtain speed information at each point of ego vehicle trajectory candidates; calculates the cost of each track through weighted summation; select the best n trajectories for the next iteration; and after a certain number of iterations; selects the one with minimum cost as the final output trajectory.
• Figure 3: The planning trajectories sampled around the initial guess are encoded as the query. The encoded map representation is reshaped into D embeddings with the same dimension as the query. The multi-head cross-attention module then outputs 2-dimensional vectors representing the recommended velocity at each query point, effectively guiding the trajectory optimization process.